SCALE: Self-Improving Web Agent via Cognitive-Aware Exploration
A team of researchers has introduced SCALE (Self-Cognitive-Aware Learning and Exploration), a framework designed for web agents that incorporates three adversarial roles: Selector, Predictor, and Judger. This framework enables autonomous identification of constraints and the broadening of cognitive limits through exploration of the environment. The strategy known as SCALE-Hop aids in global planning and helps agents steer clear of local exploration pitfalls. To enhance learning, the researchers developed SCALE-20k, a comprehensive dataset sourced from 19 real-world websites, featuring various task types and structured demonstrations derived from SCALE's exploration data. This study seeks to overcome the drawbacks of current web agents that depend on manually crafted execution processes or costly expert trajectories, with the goal of enhancing adaptability in complex, dynamic settings. The paper can be found on arXiv with the identifier 2605.31365.
Key facts
- SCALE uses three adversarial roles: Selector, Predictor, and Judger.
- SCALE-Hop is a graph exploration strategy for global planning.
- SCALE-20k dataset collected from 19 real-world websites.
- The framework aims to improve adaptability to complex environments.
- Existing web agents rely on handcrafted pipelines or expert trajectories.
- SCALE autonomously discovers agent limitations.
- SCALE expands cognitive boundaries through exploration.
- The paper is on arXiv with ID 2605.31365.
Entities
Institutions
- arXiv